US gov memo on “adversarial distillation” - are we heading toward tighter controls on open models?

A newly surfaced memo from the US Office of Science and Technology Policy (OSTP) highlights growing concerns over adversarial distillation — a method where bad actors use multiple proxy accounts and jailbreak prompts to systematically extract capabilities from frontier AI models. The memo describes this as an industrialized form of model theft, potentially enabling adversaries to replicate proprietary systems without authorization. While the immediate focus is on protecting closed, commercial models like those from OpenAI and Anthropic, the broader implications for open-weight models are significant.

The memo signals that US policymakers may begin treating model weights and advanced capabilities as strategic national security assets. This could lead to tighter export controls, licensing requirements, or even restrictions on public releases of highly capable open models. The AI community faces a pivotal question: can we preserve the innovation and accessibility of open models while addressing legitimate security concerns? The answer will shape the next phase of AI governance, balancing openness with protection against adversarial extraction at scale.